Built for the hackers. A custom Agent OS heavily inspired by OpenClaw. Honestly, there's no grand vision here—I just wanted to build a hackable Agent OS from scratch to truly figure out how this stuff works under the hood.
A personal AI agent that lives on your machine.
It uses your tools, works in your projects, and delivers real files — not just chat.
MOZI is a desktop AI agent (think of a personal, self-hosted Codex) that runs entirely on your machine. You point it at a project folder, give it a task, and it executes: shell commands, file edits, web research, document generation. Every deliverable it claims is verified against the filesystem before it reports done — no fake success.
The composer is the cockpit: pick a project (any folder or git repo), the git branch to work on, the permission level (read-only → full access), and the model — then describe what you want.
- Code — reads real repos, writes and edits files, runs tests, works on the branch you pick. A built-in branch switcher does honest
git switch(never auto-stash, never force; conflicts abort with git's own message). - Documents — generates Word / PowerPoint / Excel / PDF files and previews them inside the app: docx via a high-fidelity embedded viewer, spreadsheets as interactive grids, slides and PDFs with full CJK text. Hook up a local ONLYOFFICE container and the preview upgrades to a full editor — optional, everything degrades gracefully without it.
- Research — searches the web, reads pages and files, and compiles findings into structured reports you can open as artifacts.
- Remember — long-term memory across sessions. Tell it something once; it's there next week.
- Automate — scheduled and recurring tasks with a dedicated UI, plus reusable task templates.
- Skills — 25 built-in skills (the Anthropic official skill catalog adapted to MOZI, plus MOZI's own). Skills load on demand: the model sees a one-line catalog and pulls full instructions only when a task needs them. Drop a
SKILL.mdinto your workspace to add your own.
The desktop app is the primary way to use MOZI. Build it from source (macOS, Apple Silicon):
git clone https://github.com/spytensor/OpenMozi.git
cd OpenMozi
pnpm install
pnpm desktop:pack:mac
# → desktop/dist/mac-arm64/MOZI.app (drag into /Applications)On first launch, create your local account and add an LLM API key. The app manages its own backend and data — no separate server to run. See docs/DESKTOP-APP.md for details.
Requirements: Node.js ≥ 22 and pnpm to build. Optional extras: LibreOffice (slide/PDF conversion for previews), Docker (ONLYOFFICE editor-grade office viewing).
MOZI also runs headless with a Web UI — same runtime, same features:
pnpm mozi onboard # interactive setup: provider, API key
pnpm start # Web UI at http://localhost:9210Server-mode configuration lives in ~/.mozi/mozi.json (JSON, not YAML). Inspect or change it with pnpm mozi config get brain / pnpm mozi config set brain.model <model>, or re-run pnpm mozi onboard --update.
MOZI works with any OpenAI-compatible API — 26 providers are in the catalog. Switch anytime without losing data or history.
| Provider | Setup | Notes |
|---|---|---|
| MiniMax | MINIMAX_API_KEY |
Default provider (MiniMax-M3) |
| OpenAI | OPENAI_API_KEY |
GPT series |
| Anthropic | ANTHROPIC_API_KEY |
Claude series |
| Google Gemini | GEMINI_API_KEY |
Gemini Pro/Flash |
| DeepSeek | DEEPSEEK_API_KEY |
DeepSeek V-series, R-series |
| Moonshot | MOONSHOT_API_KEY |
Kimi models, long context |
| Groq | GROQ_API_KEY |
Ultra-fast inference |
| Ollama | Local install | Fully local and private |
| …and 18 more (xAI, Mistral, Together, OpenRouter, NVIDIA, Bedrock, …) |
Regional endpoints are supported via <PROVIDER>_BASE_URL overrides.
You (Desktop app / Web UI)
--> Gateway (sessions, auth, permission levels)
--> Brain (LLM reasoning, planning, tool calls)
--> Capabilities (shell, files, search, skills, sub-agents)
The LLM is the decision-maker; everything else is infrastructure that executes its decisions and reports the truth back. Deep dives:
pnpm dev # backend watch mode
pnpm ui:dev # Web UI dev server (hot reload)
pnpm test # full test suite (vitest, real API calls)
pnpm desktop:dev # desktop app in dev modesrc/
core/ # Brain: LLM loop, model routing, provider failover
gateway/ # sessions, auth, permission levels
capabilities/ # shell, filesystem, search, vision
skills/ # skill registry + 25 bundled SKILL.md assets
memory/ # long-term memory, vector store
store/ # SQLite (better-sqlite3)
ui/ # React + Vite desktop/web UI
desktop/ # Electron shell + packaging
Tech stack: TypeScript, Node.js 22, Fastify, better-sqlite3, Vercel AI SDK, React + Vite, Electron, Vitest.
Read CONTRIBUTING.md before opening an issue or pull request. Use the structured GitHub forms, report vulnerabilities privately, and never attach credentials or private data.
- Read
AGENTS.md,CLAUDE.md, anddocs/CONSTITUTION.mdfor repository rules. - Run relevant tests and
pnpm verify:public-exportbefore submitting changes. - Commit convention:
feat:/fix:/refactor:/docs:/test:/chore:.
GitHub Actions are intentionally disabled. Releases are built and verified locally, then uploaded to GitHub Releases with DMG, ZIP, SHA-256 checksums, and a release manifest. See docs/RELEASE.md.
Deepest respect to the OpenClaw project, which served as the primary architectural inspiration for OpenMozi.
OpenMozi's own source code is licensed under the MIT License.
OpenMozi also depends on third-party software under its own terms. In particular, the optional interactive React preview includes CodeSandbox Nodebox, which uses the Sustainable Use License and restricts commercial use and distribution. See THIRD_PARTY_NOTICES.md and the bundled license texts before redistributing OpenMozi or using that preview commercially.
MOZI — an AI that works for you, not the other way around.